Download Microarray data analysis - Ariez Medical Publishing

Practice Guidelines
Microarray data analysis: strategies,
pitfalls and applications in clinical
oncology
S.J. Van Laere, P.A. van Dam, L.Y. Dirix, P.B. Vermeulen
Microarray technology has rapidly established a critical site in cancer research. Although the
procedures to perform microarray experiments require specialised laboratory equipment and
trained personnel, the task of data analysis is often a more tantalising job. Great care must be
taken with the interpretation of the results, as there is often no a priori way to detect flaws in
the analytical procedures. The power of the microarray technology lies in its great versatility.
The technology can be used to identify new, previously undefined subgroups of supposedly
homogeneous tumour samples, to identify new (single-gene and multi-gene) biomarkers for
disease progression (prognosis) or treatment response (prediction) and to functionally and
molecularly unravel poorly characterised conditions. The purpose of this review is to provide
the reader with a glimpse of the possible strategies of analysis that exist for microarray data,
their advantages and disadvantages, their applicability to cancer research and how they might,
reshape the field of clinical oncology in the future.
(Belg J Med Oncol 2011;5:50-6)
Introduction
Since the landmark paper published by Schena et al,
describing a novel technique that allows the investigation of the expression of multiple genes in one experiment, a true avalanche of studies reporting on the
microarray technology arose.1 It took only 1 year before
the first paper applying microarrays to cancer (i.e.
melanoma) was published.2 To date, approximately
13,000 studies have used a derivate of the microarray
technology to study the biology of cancer from different perspectives (biology, diagnosis, prognosis, treatment prediction). Figure 1 demonstrates the explosive
increase in studies associating microarrays and cancer
since the inception of the technology back in 1995.
In 2003, microarrays were introduced in clinical trials.
DePrimo et al applied the technology to for the identification of predictive biomarkers for response in patients
with metastatic colorectal cancer treated with SU5416
in a Phase III trial (Figure 1).3 More recently, 2 clinical
randomised trials have been initiated to investigate the
possibility of using gene expression assays for the selection of patients with breast cancer able to benefit from
chemotherapy. The MINDACT trial aims at prospectively validating the use of the 70-gene poor prognosis
signature as a tool for the improvement of the selection
of patients with good prognosis who would not benefit from adjuvant chemotherapy.4-6 The TAILORx-trial
aims at the validation of the OncotypeDX recurrence
Authors: S.J. Van Laere MsC PhD, Oncology Centre, St. Augustinus Hospital; P.A. van Dam MD PhD; L.Y. Dirix MD PhD;
P.B. Vermeulen MD PhD, Translational Cancer Research Group (Laboratory of Pathology, University of Antwerp, Belgium; Oncology Centre,
St. Augustinus Hospital, Wilrijk, Belgium).
Please send all correspondence to: S.J. Van Laere, MsC PhD, Oncology Centre, St. Augustinus Hospital, Oosterveldlaan 24, 2610 Wilrijk,
Belgium; E-mail: [email protected].
Conflict of interest: the authors have nothing to disclose and indicate no potential conflicts of interest.
Key words: microarray, class discovery, class prediction, class comparison.
Belgian Journal of Medical Oncology
volume 5, issue 2, 2011
2
50
Practice Guidelines
score as a tool for discriminating between patients who
would benefit and patients who would not benefit from
the addition of chemotherapy to hormonal treatment.7
The principle of the microarray technology is based
on the unique feature of RNA- and DNA-molecules
to bind to RNA- or DNA molecules with a complementary sequence (i.e A with T/U and G with C).
This process is called hybridisation. When performing a microarray experiment, RNA is extracted from
a sample of interest (e.g. a breast biopsy). This RNA
sample, often referred to as the target, is subsequently converted to copy DNA (cDNA), which is further
processed and ultimately coupled to a fluorescent
dye. Once the cDNA is fluorescently labeled, it is
hybridised to the microarray. The latter is a matrix
of spots and each spot contains cDNA molecules,
often called probes, each with a unique sequence
designed to specifically bind to the cDNA molecules
from only 1 corresponding gene. During hybridisation, the cDNA molecules in the spots will capture
their fluorescently labeled complements from the
target, and, as such, the spot will acquire a fluorescent signal. The intensity of this fluorescent signal
is proportional to the amount of complementary
cDNA molecules in the target, and hence is a measure of the expression of the corresponding gene.
The microarray is scanned with an automated confocal laser scanner allowing the researcher to determine the fluorescent intensity of all spots on the
microarray. After scanning, the resulting images are
processed during the image analysis procedure to
extract raw signal intensity values. These raw signal
intensity values need to be normalised to reduce
technical variation and increase array comparability.
The guiding principle behind data normalisation is
that the majority of the genes are not differentially
expressed and, hence, that global differences in fluorescence intensities between different arrays are not
expected. After normalisation of the fluorescence
intensities, quality assessment is performed by filtering out spots with low signal-to-noise ratios, to prevent biased results.
The procedures to perform a microarray experiment
require trained personnel and specialised laboratory equipment. Nevertheless, although the process
of data analysis is logistically more feasible (it only
requires powerful computers with specific software),
it is without doubt one of the most tantalising tasks
due to the need to process vast amounts of data. To
Belgian Journal of Medical Oncology
51
Figure 1. A PubMed-search for “Microarrays and Cancer”
(blue), “Gene Signatures and Cancer” (red) and “Microarray
and Clinical Trial” (yellow) was performed for each two-year period from 1995 until 2010 (X-axis). Since the inception of the
microarray technology in 1995, the number of publications applying the technology to cancer (Y-axis) increased exponentially. Since 2005, a plateau-phase has been reached. The
number of studies reporting on gene signatures in cancer is still
exponentially increasing at the time of writing. The number of
clinical trials involving a microarray study in some way, is lagging
behind and is currently less than 5% of the total amount of
studies reporting on cancer and involving microarrays.
unravel the biological data provided by a microarray
experiment requires both skills and cautiousness,
as frequently there is no a priori way to detect flaws
in the analysis procedures. The focus of the current
review is to give the reader a glimpse of the pitfalls
associated with microarray data analysis.
Data analysis strategies
In general, 2 major analysis strategies exist; the
unsupervised and the supervised method. An
overview is presented in Table 1. The unsupervised
methods refer to a group of techniques analysing the
data without taking phenotypic features associated
with the samples into account. In an unsupervised
analysis, one is generally interested in the biological
relationship between samples. As such, this strategy is suited to discover new biological subgroups
within a set of (tumour) samples and therefore unsu-
volume 5, issue 2, 2011
Table 1. Overview of microarray data analysis strategies, the associated algorithms and their
appropriate references for further reading
STRATEGY
Unsupervised analysis
Supervised analysis
SUBSTRATEGY
ALGORITHMS
REFERENCES
Class discovery
Hierarchical cluster analysis, K-means clustering, Selforganising maps, principal component analysis
12, 13, 14, 15, 16, 17
Class comparison
T-, ANOVA-, Mann-Whitney U- and Wilcoxon Signed
Rank-testing, Linear models, Significance analysis of
microarrays, Permutation testing
12, 18,19, 20, 21
Class prediction
Nearest centroid classification, Neural networks, Linear
discriminant analysis, K-nearest neighbours, Support
vector machines
15, 22, 23, 24, 25
Functional analysis
Gene Set Enrichment Analysis, Global testing, Global AN26, 27, 28, 29, 30
COVA models, Pseudo-comparative genomic hybridisation
pervised analysis strategies are often referred to as
class discovery methods. For example, in a study by
Perou et al, the authors performed an unsupervised
analysis on a set of breast tumour samples.8 This
resulted in 2 major subgroups associated with the
presence or absence of Estrogen Receptor (ER) protein expression. In addition, new subgroups could
be identified in both sets. The authors concluded
that breast cancer is composed of different subtypes,
some of which are biologically more related (e.g the
ER+ subtype Luminal A and B) than others.8 This
example demonstrated that class discovery analysis
assists in identifying novel tumour subclasses and
thus paves the way for more personalised treatment.
The supervised methods differ from the unsupervised
methods in that the associated phenotypic data of the
tumour samples are integrated in the analysis. The
supervised methods can be subdivided in class comparison- and class prediction-methods. As the name
suggests, class comparison-methods essentially compare 2 or more classes or groups of tumour samples,
mainly to identify differentially expressed genes. This
method is well-suited to identify individual biomarkers, but can also be used as a starting point for downstream functional analysis (vide infra) For example,
our research group used microarray analysis to molecularly characterise inflammatory breast cancer (IBC).
Therefore, we compared a group of IBC samples to a
group of nIBC samples.9 In order to be able to compare both groups, one needs a priori knowledge of the
sample or grouping labels and consequently this is a
supervised analysis strategy.
The second supervised analysis strategy, class prediction, refers to a set of techniques that is used to
design gene signatures. A gene signature is a bio-
marker consisting of a collection of genes selected
for their ability to discriminate with high specificity
and sensitivity between 2 or more groups of tumour
samples. In essence, each gene comprising a gene
signature is a biomarker on its own. For the gene
selection procedure, again, this technique requires
prior knowledge of the sample or group labels.
Gene signatures are among the holy grails in cancer research due to their clinical potential, as they
can be used both as prognostic and therapeutic (i.e.
predictive) markers. For example, Van ‘t Veer et al
described a gene signature composed of 70 genes
able to discriminate between patients with lymph
node negative breast cancer who developed distant
metastases within 5 years and those who remained
metastases-free for more than 5 years.10 An example of a predictive study is described by Hess et al,
who designed a 30 gene signature capable of predicting pathological complete response to neoadjuvant Paclitaxel and Fluorouracil + Doxorubicin +
Cyclophosphamide in patients with breast cancer.11
At the time this review was written, about 2,300
papers associating gene signatures with cancer have
been published (Figure 1).
From the perspective of tailored treatment, the development of gene signatures allowing for the identification of samples with certain activated pathways is
particularly interesting. For example, Creighton et al
designed a gene signature predicting the activation
of the IGF1-induced signal transduction pathway in
breast cancer. They demonstrated that this gene signature is able to predict sensitivity to anti-IGF1R therapy in cell lines and xenografts. Particularly cell lines
and xenografts of triple negative breast tumours were
predicted to be highly sensitive to anti-IGF1 therapy
2
52
Practice Guidelines
(BMS-754807). Indeed, treatment of xenografts or
triple negative breast tumours with BMS-754807 in
combination with Docetaxel demonstrated significant
tumour regression until no tumour was palpable.32
In addition, the same authors demonstrated that a
subgroup of ER+ breast tumours, particularly the
endocrine resistant Luminal B tumours, exhibit the
expression of the IGF1-activated gene signature.33
Combined, these data suggest that patients with
triple negative or endocrine resistant breast cancer
could benefit from anti-IGF1R therapy. More importantly, these data also demonstrate how prediction of
pathway activation through gene expression profiling
can become a valuable tool in establishing tailored
treatment.
A third analysis strategy that needs mentioning is
the functional analysis. This too can be regarded
as a supervised analysis strategy, because its starting point is either the class prediction or the class
comparison analysis. The functional analysis translates lists of genes into biological processes or signal transduction pathways. The rationale behind the
functional analysis is, that most genes can play vital
roles in different processes or signal transduction
pathways. Therefore, it is difficult to unravel the biology based upon lists of individual genes alone. On
the other hand, a biological process or signal transduction pathway is usually composed of a unique
collection of genes. If these process- or pathway-specific genes can be identified in a list of differentially
expressed genes, the associated biological process or
signal transduction pathway will be of importance
to the biology of the samples that have been used
to define the list of differentially expressed genes.
In addition, the functional analysis has the advantage of downscaling the amount of possibly involved
parameters to a more feasible number. The gene
lists associated with biological processes and signal transduction pathways can be found in publicly
available databases like Gene Ontology (GO), Kyoto
Encyclopedia of Genes and Genomes (KEGG) and
Transcription Factor databases (TransFac).27
In the remaining parts of this review, the aspects of the
above outlined analysis strategies (class discovery, class
comparison and class prediction) will be discussed
in more detail with special focus on potential pitfalls.
For more elaborate background on the items and algorithms discussed in the following sections, the reader is
directed towards the references in Table 1.
Belgian Journal of Medical Oncology
53
Class discovery analysis
Class discovery analysis is an unsupervised analysis directed at analysing the biological relationships
between the samples. The biological relationships
between samples are defined by how similar the
gene expression profiles of the samples are. This
implies that the algorithms need a measure of similarity. The most widely used similarity metrics are
the Euclidean distance, the correlation coefficient
and the Manhattan distance, but many more exist.
Class discovery algorithms organise data in such
way that samples with similar expression profiles
are grouped together and that the gene expression
profiles of the samples in different groups are maximally dissimilar. Groups of samples resulting from
class discovery algorithms are called clusters and the
algorithms themselves are often referred to as cluster
algorithms. Examples of these algorithms are hierarchical cluster analysis, k-means cluster analysis and
self-organising maps.12-17
Another powerful class discovery algorithm that is
widely gaining more and more interest is principal
component analysis (PCA). Although this algorithm
is regarded as an unsupervised analysis method, it
should be mentioned separately from the cluster
algorithms because no real distance metric is defined
from the start. The rationale of PCA is to visualise
multidimensional data in 2 or 3 dimensions through
data decomposition. An average microarray experiment consists of 100 samples and 20,000 genes.
This means that the 100 samples can be visualised
in a 20,000-dimensional hyperspace. PCA can be
used to visualise these multidimensional datasets by
reducing the number of dimensions down to 2 or
3 theoretical dimensions, the principal components.
These capture most of the gene expression variation
present in the dataset and, as such, preserve the gene
expression relationships between the samples.12
Although class discovery techniques are extremely
powerful, great care must be taken in applying these
algorithms. Even though the methods used are
objective in the sense that the algorithms are welldefined and reproducible, they are still subjective in
that selecting different algorithms or different distance metrics will place different objects into different clusters. Furthermore, clustering unrelated data
would still produce clusters, although they might
not be biologically meaningful. The challenge is
therefore, to select the data and to apply the algo-
volume 5, issue 2, 2011
rithms appropriately so that the classification is sensible.12 To obtain a useful analysis while at the same
time preserving the expression relationships in the
array dataset, it is standard practice to use those
genes with a high coefficient of variation and mean
expression level. In addition, each clustering result
should be analysed for cluster stability or robustness
to prevent reporting results that are not supported
by the data.
Class comparison analysis
In a class comparison analysis, one is interested in
finding lists of genes that are differentially expressed
between the conditions during the study. These gene
lists can be used to unravel the molecular biology of
previously uncharacterised conditions or they can be
used to define potential biomarkers to follow-up on
treatment or disease progression. Different methods
exist to define lists of differentially expressed genes,
all derived from the field of traditional statistics, for
example T-testing, ANOVA-testing, Mann-Whitney
U-testing and Wilcoxon Signed Rank-testing. Also
linear models (i.e. linear regression analysis) can be
used for this purpose. For example, when studying
differential gene expression between ER+ and ERtumours, the power of each gene in the microarray
experiment to correctly predict ER+ from ER- breast
tumours is evaluated using linear regression analysis. In this, the predictive power is proportional to
the level of differential expression.
The greatest statistical challenge is based on the
conclusions concerning the expression of a great
number of genes on the basis of a small number
of samples. This problem is also referred to as the
multiple testing or multiple comparisons problem.
The major culprit for this phenomenon is the significance level or alpha-level that is the preset probability that a statistical test will be a false positive.
Another way to consider the significance level is
that it is the percentage of statistical tests that will
be false positive. In microarray analysis, which usually deals with on average 20,000 genes, this means
that 1,000 false positives are expected for a significance level of 0.05. For this kind of experiment, the
False Discovery Rate (FDR), which is the percentage
of false positives in the total amount of significant
results, can easily be about 50%. The problem is that
one cannot determine which statistical results are
Belgian Journal of Medical Oncology
true/false positives.12,18-21. A number of algorithms
have been developed to tackle this problem. The
Bonferroni correction is very stringent and aims at
reducing the probability of having only one false
positive in the full set of comparisons. This method
is widely used when dealing with the identification
of individual biomarkers. The FDR-correction proposed by Benjamini and Hochberg aims at reducing the expected proportion of false positives.20 This
method is less stringent and is more appropriate
when aiming at the molecular characterization of a
condition or the identification of gene signatures.
FDR-levels of up to 10% are considered appropriate
in microarray literature.12, 18-21 Of note, biologically
interesting genes with more elevated FDR-levels (i.e.
greater than 10%) can still be relevant. For example,
genes with an FDR of 20% still have 80% chance of
being true positives. It is standard practice to evaluate such genes with alternative expression profiling
techniques (e.g. qRT-PCR), preferentially on a different and larger set of samples.
Class comparison analysis
Perhaps the most promising application of microarrays for expression profiling, is class prediction.
In this setting it is not necessary to understand
the underlying molecular biology of the condition.
Rather, it is a purely statistical exercise in linking a
certain pattern of expression to a certain diagnosis
or prognosis.15,22-25 The procedure of classification
is a well-established field in statistics, from which a
wealth of methods can be drawn. Each modelling
method consists of 3 stages: feature selection, model
building and model assessment.15,22-25
Feature selection involves the process of selecting genes that will be used to construct the classifier. Feature selection should favour informative
genes without being too restrictive in their selection
criteria. Usually, simple fold change statistics, T-test
or ANOVA-statistics (for multiple groups) are used.
Also, genes can be selected based on their standard
deviation. Genes that ideally compose a gene signature are those genes demonstrating a huge difference
in mean or median gene expression between groups of
tumour samples (e.g. high fold-change or high T-test
or ANOVA-statistics) but show little variation within
each group of tumour samples under study (e.g. low
within-group standard deviation). After the feature
volume 5, issue 2, 2011
2
54
Practice Guidelines
Key messages for clinical practice
1.
Class discovery analysis can be used to identify novel tumour subgroups and, as
such, can assist in paving the way for tailored therapy.
2.
Class comparison analysis can be used to study the biology of poorly characterised
tumour specimens and to identify potential molecular targets for targeted therapy.
3.
Class prediction analysis can be used to identify multi-gene biomarkers to assist
in therapeutic decision-making and patient prognosis.
4.
Gene expression profiling can provide guidance in the establishment of tailored
therapy through prediction of signal transduction pathway activation in clinical
tumour samples.
selection process, the model should be built. Many
model-building procedures exist, including nearest
centroid classification, neural networks, linear discriminant analysis, K-nearest neighbours and support vector machines. A popular choice amongst these is the
nearest centroid classification, which compares each
sample under study to a set of centroids, one for each
class. A centroid is the average expression profile of a
certain class and is based on the genes selected during
the feature selection process. The final step in model
building is to define a prediction rule, for example a
threshold on the Euclidean distance between the samples and the centroids. When this Euclidean distance
exceeds a given value, the sample is not classified in
the group represented by the centroid and vice versa.
Once the prediction model is constructed, the next
step is model assessment and validation, essentially
performed on a series of samples not used for feature
selection or model-building. In the model assessment
stage, metrics such as sensitivity (% of correctly classified positive samples), specificity (% of correctly classified negative samples), accuracy (% correct classifications) and prediction error (% misclassifications) can
be examined to determine how well the prediction
model is able to correctly classify a series of unknown
samples. For example, we designed a gene signature
able to predict IBC samples from nIBC samples.
Testing this gene signature onto an independent series
of IBC/nIBC samples proved that the gene signature
correctly classified all IBC samples (sensitivity=100%)
and only 1 nIBC sample was misclassified (specific-
Belgian Journal of Medical Oncology
55
ity=91%). The total percentage of misclassifications
was 5.8% (=prediction error) and the accuracy was
94.2%.31 The rationale behind model assessment is,
that a classifier can perform well on the samples used
for feature selection/model-building, but may perform
poor on other samples subject to the same classification procedure. This phenomenon is called overfitting.
The best way to perform model assessment is to use
an independent dataset, composed of samples not
used in the feature selection/model-building phase.
This can be done by splitting a group of samples in a
training set, used for feature selection/model-building
and a test set used for model assessment. If the initial
set of samples is too small for splitting, other model
assessment procedures exist such as cross-validation
or bootstrap analysis.15,22-25
Conclusion
Since the introduction of the microarray technology
in 1996, the technique has acquired a permanent status in cancer research due to its versatile applicability.
The technique can be used to discover new tumour
subtypes, to identify single-gene or multiple-gene biomarkers for response to treatment or disease progression and to unravel the molecular biology of previously
uncharacterised or poorly characterised conditions.
Nevertheless, despite the wide range of possibilities,
great care must be taken when analysing data, as many
pitfalls exist that can easily trap a researcher into reporting results unsupported by the data. In addition, as
volume 5, issue 2, 2011
microarray technology is gradually finding its way into
clinical trials, knowledge of the possible analysis strategies and existing methods, with their advantages and
disadvantages, is important to the clinician in order to
be able to interpret the results correctly.
15. Boutros PC, Okey AB. Unsupervised pattern recognition: an introduction to the
whys and wherefores of clustering microarray data. Brief Bioinform 2005;6:331-43.
16. Eisen MB, Spellman PT, Brown PO, Botstein. Cluster analysis and display of
genome-wide expression patterns. Proc Natl Acad Sci U S A 1998;95:14863-8.
17. Azuaje F. Clustering-based approaches to discovering and visualising
microarray data patterns. Brief Bioinform 2003;4:31-42.
18. Reiner A, Yekutieli D, Benjamini Y. Identifying differentially expressed genes
References
using false discovery rate controlling procedures. Bioinformatics 2003;19:368-75.
1. Schena M, Shalon D, Davis RW, Brown PO. Quantitative monitoring of
19. Dudoit S. Statistical methods for identifying differentially expressed genes
gene expression patterns with a complementary DNA microarray. Science
in replicated cDNA micorarray experiments. Stat Sinica 2002;12:111-39.
1995;270:467-70.
20. Benjamini Y. Controlling the false discovery rate: a practical and powerful
2. DeRisi J, Penland L, Brown PO, Bittner ML, Meltzer PS, Ray M, et al. Use
approach to multiple testing. J Roy Stat Soc B 1995;57:289-300.
of a cDNA microarray to analyse gene expression patterns in human cancer.
21. Tusher VG, Tibshirani R, Chu G. Significance analysis of microarrays applied
Nat Genet 1996;14:457-60.
to the ionizing radiation response. Proc Natl Acad Sci U S A 2001;98:5116-21.
3. DePrimo SE, Wong LM, Khatry DB, Nicholas SL, Manning WC, Smolich
22. Simon R. Roadmap for developing and validating therapeutically relevant
BD, et al. Expression profiling of blood samples from an SU5416 Phase III
genomic classifiers. J Clin Oncol 2005;23:7332-41.
metastatic colorectal cancer clinical trial: a novel strategy for biomarker identi-
23. Huang J, Bhandarkar SM. A comparison of physical mapping algorithms
fication. BMC Cancer 2003;3:3.
based on the maximum likelihood model. Bioinformatics 2003;19:1303-10.
4. Cardoso F, Van’t Veer L, Rutgers E, Loi S, Mook S, Piccart-Gebhart MJ. Clinical
24. Pirooznia M, Yang JY, Yang MQ, Deng Y. A comparative study of
application of the 70-gene profile: the MINDACT trial. J Clin Oncol 2008;26:729-35.
different machine learning methods on microarray gene expression data.
5. Cardoso F, Piccart-Gebhart M, Van’t Veer L, Rutgers E. The MINDACT trial: the
BMC Genomics 2008;9:S13.
first prospective clinical validation of a genomic tool. Mol Oncol 2007;1:246-51.
25. Ambroise C, McLachlan GJ. Selection bias in gene extraction on the basis
6. Mook S, Van’t Veer LJ, Rutgers EJ, Piccart-Gebhart MJ, Cardoso F.
of microarray gene-expression data. Proc Natl Acad Sci U S A 2002;99:6562-6.
Individualization of therapy using Mammaprint: from development to the
26. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA,
MINDACT Trial. Cancer Genomics Proteomics 2007;4:147-55.
et al. Gene set enrichment analysis: a knowledge-based approach for interpreting
7. Sotiriou C, Piccart MJ. Taking gene-expression profiling to the clinic: when
genome-wide expression profiles. Proc Natl Acad Sci U S A 2005;102:15545-50.
will molecular signatures become relevant to patient care? Nat Rev Cancer
27. Stoeckert CJ Jr, Causton HC, Ball CA. Microarray databases: standards
2007;7:545-53.
and ontologies. Nat Genet 2002;32 Suppl:469-73.
8. Perou CM, Sørlie T, Eisen MB, Van de Rijn M, Jeffrey SS, Rees CA, et al.
28. Furge KA, Dykema KJ, Ho C, Chen X. Comparison of array-based com-
Molecular portraits of human breast tumours. Nature 2000;406:747-52.
parative genomic hybridization with gene expression-based regional expres-
9. Van Laere S, Van der Auwera I, Van den Eynden G, Van Hummelen P,
sion biases to identify genetic abnormalities in hepatocellular carcinoma.
Van Dam P, Van Marck E, et al. Distinct molecular phenotype of inflammatory
BMC Genomics 2005;6:67.
breast cancer compared to non-inflammatory breast cancer using Affymetrix-
29. Goeman JJ, Bühlmann P. Analyzing gene expression data in terms of
based genome-wide gene-expression analysis. Br J Cancer 2007;97:1165-74.
gene sets: methodological issues. Bioinformatics 2007;23:980-7.
10. Van ‘t Veer LJ, Dai H, Van de Vijver MJ, He YD, Hart AA, Mao M, et al.
30. Goeman JJ, Van de Geer SA, De Kort F, Van Houwelingen HC. A
Gene expression profiling predicts clinical outcome of breast cancer. Nature
global test for groups of genes: testing association with a clinical outcome.
2002;415:530-6.
Bioinformatics 2004;20:93-9.
11. Hess KR, Anderson K, Symmans WF, Valero V, Ibrahim N, Mejia JA,
31. Van Laere SJ, Beissbarth T, Van der Auwera I, Van den Eynden GG,
et al. Pharmacogenomic predictor of sensitivity to preoperative chemotherapy
Trinh XB, Elst H, et al. Relapse-free survival in breast cancer patients is asso-
with paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide in breast
ciated with a gene expression signature characteristic for inflammatory breast
cancer. J Clin Oncol 2006;24:4236-44.
cancer. Clin Cancer Res 2008;14:7452-60.
12. Quackenbush. Computational analysis of microarray data. Nat Rev Genet
32. Litzenburger BC, Creighton CJ, Tsimelzon A, Chan BT, Hilsenbeck SG,
2001;2:418-27.
Wang T, et al. High IGF-IR activity in triple-negative breast cancer cell lines and
13. Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP,
tumorgrafts correlates with sensitivity to anti-IGF-IR therapy. Clin Cancer Res
et al. Molecular classification of cancer: class discovery and class prediction
December 2010 [Epub ahead of print]
by gene expression monitoring. Science 1999;286:531-7.
33. Creighton CJ, Casa A, Lazard Z, Huang S, Tsimelzon A, Hilsenbeck SG,
14. Clarke R, Ressom HW, Wang A, Xuan J, Liu MC, Gehan EA, et al. The
et al. Insulin-like growth factor-I activates gene transcription programs
properties of high-dimensional data spaces: implications for exploring gene
strongly associated with poor breast cancer prognosis. J Clin Oncol
and protein expression data. Nat Rev Cancer 2008;8:37-49.
2008;26:4078-85.
Belgian Journal of Medical Oncology
volume 5, issue 2, 2011
56
2